Enabling Deep-Neural-Network-Integrated Optical and SAR Data to Estimate the Maize Leaf Area Index and Biomass with Limited In Situ Data

Round 1

Reviewer 1 Report

This manuscript presented a study about enabling deep neural network integrated optical and SAR data for estimating maize LAI and biomass with limited In-situ data. The study is interesting for the topic of Remote Sensing. There are some comments to improve the manuscript.

1. The Abstract should be shorten to clear express the mainly research gaps of current study and the contribution of this study.

2. Formula (14) and (15) should be deleted, these two indicators are commonly used for accuracy assessment.

3. Section 3.1 and section 3.2, there are many descriptions about methods instead of results.

4. A figure about the spatial distribution of the estimated LAI and biomass is needed.

5. Conclusions part is mainly the repetition of results, it should be shorten to clearly show the findings of this study.

Author Response

Dear Reviewer,

Thank you very much for your interest in our work, and we really appreciate your good suggestions to help us improve the paper.  Please check our point-by-point response to the comments (below):

Reviewer#1, Concern # 1: The Abstract should be shorten to clear express the mainly research gaps of current study and the contribution of this study.

Author response: Thank you very much for your good suggestion. We have shorten the abstract as you suggested.

Author action: We updated the manuscript by deleting some detail information and make the abstract much more clear.

 

Reviewer#1, Concern # 2: Formula (14) and (15) should be deleted, these two indicators are commonly used for accuracy assessment.

Author response: Thank you very much for your good suggestion. We have deleted Formula (14) and (15).

Author action: We updated the manuscript by deleting Formula (14) and (15).

 

Reviewer#1, Concern # 3: Section 3.1 and section 3.2, there are many descriptions about methods instead of results.

Author response:  Thank you very much for your good suggestion. We had modified it .

Author action: We updated the manuscript by moving Section 3.1 and section 3.2 to the former Section {Method}.

 

Reviewer#1, Concern # 4: A figure about the spatial distribution of the estimated LAI and biomass is needed.

Author response:  Thank you very much for your good suggestion. We have corrected it as you suggested.

Author action: We updated the manuscript by adding Section4.3 “Results of maize LAI and biomass estimation based on…” including the figure about the spatial distribution of the estimated LAI and biomass.

 

Reviewer#1, Concern # 5: Conclusions part is mainly the repetition of results, it should be shorten to clearly show the findings of this study.

Author response:  Thank you very much for your good suggestion. We have shortened the conclusion part.

Author action: We updated the manuscript by showing findings rather than results, as shown in Section6.

We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper.

We appreciate for your warm work earnestly and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Reviewer 2 Report

Enabling the Deep Neural Network Integrated Optical and SAR Data to Estimate Maize Leaf Area Index and Biomass With Limited In-situ Data

 

Line 12 in abstract "thus integrate them to make full use". Please make it clear

The introduction section is explained very well and enough literature is cited in the section

Materials and methods has been discussed with enough details and written very well. 

Why even the basic baseline models perform better than GSDNN without mixup+? (Table 4)

What are the effects of fusing optical and SAR and how can they work individually without fusion for estimating LAI and biomass?

 

Overall the study is very nice, the paper is very well written and I recommend it for publication after minor revisions. 

Updated more comments, please see the attachment.

Comments for author File: Comments.pdf

Author Response

Dear Reviewer,

Thank you very much for your interest for our work, and we really appreciate your good suggestions help to improve our manuscript. Here are our point-by-point response to the comments (below):

Reviewer#2, Concern # 1: Leaf area index and biomass estimation using deep learning is very prevailing these days. The authors need to explain what new knowledge they are adding to the topic.

Author response:  Thank you very much for your interest in our research. We really appreciate your willingness to discuss what new knowledge using deep learning to estimate LAI and biomass adding to the topic. 

      Firstly, although deep learning is very prevailing these days, few papers focus on the question about how to address the limited in-situ dada in leaf area index and biomass estimation. As we all know, deep learning is a data-driven method, so less training data will cause under-fitting or over-fitting problems. So  bridging the gap of data shortage is necessary. Here, we proposed a method termed “mixup+”.

Secondly, although some paper used some ANN method or some simple network in leaf area index and biomass estimation, neither of them is appropriate for integrate optical and SAR data. So we design a neural network based on Siamese architecture which has two branches for optical and SAR input and fusion process.

    Moreover, we adopt the gating mechanism to help optical and SAR data to realize effective interaction during the fusion process. So this design can help us to realize the effective and deep fusion of optical and SAR data.

Author action: We have explained what new knowledge using deep learning to estimate LAI and biomass adding to the topic in Author response above.

 

Reviewer#2, Concern # 2: Abstract is too long and consist a lot of text. Please consider shortening it.  

Author response: Thank you very much for your good suggestion. We have shorten the abstract here.

Author action: We updated the manuscript by deleting some detail information and make the abstract much more clear.

 

Reviewer#2, Concern # 3: Why even the basic baseline models perform better than GSDNN without mixup+? (Table 4)

Author response:  We really appreciate your good question. That’s because GSDNN is deep learning method which is data-driven, so it tends to be over-fitting or under-fitting with limited in-situ data. So with the help of mixup+, GSDNN achieves better result in LAI and biomass estimation compared to baseline models.

Author action: We answered this question as above.

Reviewer#2, Concern # 4: What are the effects of fusing optical and SAR and how can they work individually without fusion for estimating LAI and biomass?

Author response: Thank you very much for your good question. Due to the different imaging mechanism of optical and SAR data, both of them have advantages and disadvantages in LAI and biomass estimates.  As the table 5 shows, after integrating optical and SAR data, the accuracy is improved compared to only using optical or SAR data(Take MLP as example).  In addition, because GSDNN is designed for integrating optical and SAR data, so it cannot obtain the result only using optical or SAR data.

Author action: We answered this question as above.

 

Reviewer#2, Concern # 5: For implementation, the authors need to mentioned which platform was used to train models.  there is no information about any programming language or framework where deep learning models were trained. It needs to be included briefly for better reproduction.

Author response:  Thank you very much for your good suggestion. In this study, we use PyTorch deep learning framework based on python 3.6.8. We run all experiments on GeForce RTX 2080Ti GPU and Ubuntu 18.04. We have added the detail information in Section 3.5 .

Author action: We updated the manuscript by adding “In this study, we use PyTorch deep learning framework based on python 3.6.8. We run all experiments on GeForce RTX 2080Ti GPU and Ubuntu 18.04.” in section 3.5 “Settings and training details for GSDNN”.

 

Reviewer#2, Concern # 6: What train/test splitting was used to split the data between training and testing samples, if it was done?

Author response:  Thank you very much for your question. In this study, to reduce the impact of data randomness, five-fold cross-validation was used to assess the accuracy of the LAI and biomass estimation models. Firstly, the data set was randomly divided into five parts and each of them was used as test data to train the model alternately. And then, 10% of the data from the remaining four parts was randomly selected as the verification set, and all the remaining data as the training set.

Author action: We answered the question as above.

 

Reviewer#2, Concern # 7: Figure 4 and Table 3. Please include if they are training or testing results?

Author response:  Thank you very much for your good suggestion.  Figure 4 and Table 3 are testing results.

Author action: We updated the manuscript by adding Figure 4 and Table 3 are testing results, as shown in Section4.

 

Reviewer#2, Concern # 8: Generally, training results are very higher and can lead to overfitting; in that case, the models perform very poorly for testing data. Did the authors test this?

Author response:  Thank you very much for your good suggestion. Yes, we have taken this into consideration. To avoid this problem, we use five-fold cross-validation to train and test the model.  In order to reduce the impact of randomness of training and data, this paper conducts 100 times of 5-fold cross validation, and then calculates and reports the mean and variance on the results of 500 experiments in total.

Author action: We answered the question as above.

 

Reviewer#2, Concern # 9: Same comment is for results presented in Table 4. It should be very clearly mentioned if those results are from training or testing set.

Author response:  Thank you very much for your good suggestion.  We have modified it according to your ideas.

Author action: We updated the manuscript by adding “(all of them are testing results)” in the article, as shown in Section3.1 and 3.2.

 

Reviewer#2, Concern # 10: In case of comment 7,8 and 9 it is my best guess that the results are from training set. If it is so, please also include testing results.

Author response:  Thank you very much for your good suggestion.  we have modified it as you suggested.

Author action: We updated the manuscript by adding “all results are testing results” in the paper.

 

Reviewer#2, Concern # 11:

Author response:  I am so sorry Comment 11 is empty in the system.

Author action:

 

Reviewer#2, Concern # 12: Correct the heading 0. Introduction

Author response: Thank you very much for your good suggestions. We have modified all headings according to your suggestions.

Author action: We updated the manuscript by modified “0. Introduction” to “1. Introduction”, and other headings have also been corrected accordingly.

 

Reviewer#2, Concern # 13: The number of samples seems to be very low for any deep learning model which generally works on samples higher in number. How does the author justify applications of neural networks with such limited number of samples i.e. 40.

Author response: Thank you very much for your good questions. We really appreciate your willingness to discuss the problem about limited samples. As we all know, few samples will result in over-fitting or under-fitting problems for deep learning, so the model will not perform well with limited samples. We used 158 samples in our research including 40 samples from four different growth stages respectively. Compared to other traditional machine learning methods such as MLR,PLR,RFR,SVR, our method GSDNN does not perform well based on machine learning, So the lower sample amounts like 40 will not perform well due to the over-fitting or under-fitting problems. In addition, we can compare the accuracy of the deep learning models with other traditional learning models to judge the quality of the model with limited samples.  We also can judge the quality of the model with limited samples by running more times to observe their performance.

Author action: We answered the question as above.

 

Reviewer#2, Concern # 14: For better reproducablity the authors needs to share the codes and results using some open source repositories such as github?

Author response: We really appreciate your good suggestions. We are very happy to share our codes and results with other researches.

Author action: We will share our codes and results after this paper accepted.

We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper.

We appreciate your warm work earnestly and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Reviewer 3 Report

In this paper, the authors proposed a GSDNN and optical, SAR data fusion method for maize LAI and biomass estimation. Overall, this paper looks good. Here are some detailed comments. 

1. Generally, this paper only has one-year data from a few days. Data representativeness is one of my concerns. 

2. Line 11, use the full name for "SAR" when first mentioned in the paper. 

3. Line 72, "consequently", avoid the grammar mistake

4.  Line 132, wrong format for longitude and latitude

5.  Line 166, "select", kee your past tense consistent in one paragraph

6.  Line 192 equation (3), add a comma at the end of your equation, like what you did for equation (16); equations should be considered as one part of your paragraph. Fix this problem for all your equations.

 7. Section 2.2, considering one of your key contributions to this paper is the GSDNN, you should explain why your GSDNN is better theoretically if possible. 

8. Line 254, wrong citation?

9. Figure 3 is a common workflow for any neural network, it looks good but try to add some unique descriptions for your neural network workflow. 

10. Try to add training or testing process figures for your GSDNN. 

11. Line 417, do not put "And" at the beginning of your sentence, fix this problem for your whole paper. 

12. Many researchers start to care about the reproducibility of research papers. Try to explain how others can reproduce your work. 

Author Response

Dear Reviewer,

Thank you very much for your interest for our work, and we really appreciate your good suggestions help to improve our manuscript. Here are our point-by-point response to the comments (below):

Reviewer#3, Concern # 1: Generally, this paper only has one-year data from a few days. Data representativeness is one of my concerns.

Author response: Thank you very much for your good questions. We really only has one-year data in our paper which may cause some representativeness as you mentioned. Considering this problem, we adopted the following two methods to ensure the representativeness of samples as much as possible: on the one hand, we adopted the random sampling method to collect samples, and collected 3 samples at each sample point to take the mean value; On the other hand, the samples we collected cover all the key growth stages of maize, including jointing stage, trumpet stage, flowering stage and filling stage. Please allow us to continue to collect samples of different crops for further validation of the model in future research work. Thank you very much for your understanding and support.   

Author action: we answered the question as above. We will collect more samples of different crops for further validation of the model in our future research work and we mentioned it in Section6.

 

Reviewer#3, Concern # 2: Line 11, use the full name for "SAR" when first mentioned in the paper.

Author response: Thank you very much for your good suggestion. We have corrected it as you suggested.

Author action: We updated the manuscript by using “synthetic aperture radar (SAR)” instead of “SAR”.

 

Reviewer#3, Concern # 3: Line 72, "consequently", avoid the grammar mistake.

Author response: Thank you very much for your good suggestion. We are very sorry for our careless mistake and it was corrected.

Author action: We updated the manuscript by changing “consequently” to “Consequently”.

 

Reviewer#3, Concern # 4: Line 132, wrong format for longitude and latitude.

Author response: Thank you very much for your good suggestion. We are very sorry for the wrong format here and we corrected it here.

Author action: We updated the manuscript by changing “39◦54.20N, 117◦26.90E” to “ 39°54.20N, 117°26.90E”.

 

Reviewer#3, Concern # 5: Line 166, "select", keep your past tense consistent in one paragraph.

Author response: We really appreciate your comments. We corrected it as you suggested.

Author action: We updated the manuscript by changing “select” into “selected”.

 

Reviewer#3, Concern # 6: Line 192 equation (3), add a comma at the end of your equation, like what you did for equation (16); equations should be considered as one part of your paragraph. Fix this problem for all your equations.

Author response: Thank you very much for your good suggestion. We have corrected it as you suggested.

Author action: We updated the manuscript by adding a comma at the end of equation (3).

 

Reviewer#3, Concern # 7: Section 2.2, considering one of your key contributions to this paper is the GSDNN, you should explain why your GSDNN is better theoretically if possible.

Author response: Thank you very much for your good suggestion. Thank you very much for your interest in our research. We really appreciate your willingness to discuss why GSDNN is better theoretically . 

      Firstly, although some paper used some ANN method or some simple network in leaf area index and biomass estimation, neither of them is appropriate for integrate optical and SAR data. So we design a neural network based on Siamese architecture which has two branches for optical and SAR input and fusion process, which can help to integrate optical and SAR data.

In addition, we adopt the gating mechanism to help optical and SAR data to realize effective interaction during the fusion process. So this design can help us to realize the effective and deep fusion of optical and SAR data.

All in all, GSDNN has great advantages in better integration of optical and SAR data to estimate LAI and biomass.

Author action: We updated the manuscript by deleting some detail information mentioned in the paper and make the abstract much more clear.

 

Reviewer#3, Concern # 8: Line 254, wrong citation?

Author response: Thank you very much for your good suggestion. We have corrected it here.

Author action: We updated the manuscript by correcting the citation here, thank you very much for your help to find this mistake.

 

Reviewer#3, Concern # 9: Figure 3 is a common workflow for any neural network, it looks good but try to add some unique descriptions for your neural network workflow.

Author response: Thank you very much for your good suggestion. We have modified it as you suggested.

Author action: We updated the manuscript by adding more detail information in the Figure3.

 

Reviewer#3, Concern # 10: Try to add training or testing process figures for your GSDNN.

Author response: Thank you very much for your good suggestion. We have added it as you suggested.

Author action: We updated the manuscript by adding the training or testing process figures for both LAI and  biomass in our paper, as shown in Figure 5.

 

Reviewer#3, Concern # 11: Line 417, do not put "And" at the beginning of your sentence, fix this problem for your whole paper.

Author response: We really appreciate your good suggestion. We have corrected all of this problem in our paper.

Author action: We updated the manuscript by changing “And then we proposed a novel deep neural network called GSDNN…” into “The GSDNN proposed in this study… ”. We corrected all of this problem in our paper.

 

Reviewer#3, Concern # 12: Many researchers start to care about the reproducibility of research papers. Try to explain how others can reproduce your work.

Author response: We really appreciate your good suggestion. We will share our codes and results after our paper accepted.

Author action: We answered the question as above.

We tried our best to improve the manuscript and made some changes in the manuscript. These changes will not influence the content and framework of the paper.

We appreciate your warm work earnestly and hope that the correction will meet with approval.

Once again, thank you very much for your comments and suggestions.

Round 2

Reviewer 1 Report

The authors have revised the manuscript according to the comments.